Accelerating the Discovery of Deep-Ultraviolet Nonlinear Optical Materials by Combining Machine Learning Interatomic Potentials with First-Principles Calculations: A Case of LiB2O3F

Abstract

Discovering deep-ultraviolet (deep-UV) nonlinear optical (NLO) crystals is hampered by the vast structural search space and the cost of first-principles screening. Here, we establish an integrated framework coupling machine learning interatomic potential (MLIP) construction, MLIP-assisted crystal structure prediction (CSP), and first-principles property calculations and apply it to the LiB2O3F system. A total of 40 low-energy candidate structures with formation energies above the thermodynamic convex hull (Ehull) ≤ 50 meV/atom were identified. Notably, seven of these structures are thermodynamically competitive with zero/near-zero Ehull values, namely, LiB2O3F-1 (P3, Z = 6), LiB2O3F-2 (R3c, Z = 6), LiB2O3F-3 (R3, Z = 3), LiB2O3F-4 (P3, Z = 3), LiB2O3F-5 (P31c, Z = 2), LiB2O3F-6 (P3, Z = 6), and LiB2O3F-7 (P63, Z = 2). First-principles calculations further reveal that these thermodynamically competitive phases exhibit wide band gaps ranging from 7.737 to 8.133 eV at the HSE06 level, suitable second harmonic generation (SHG) coefficient magnitudes ranging from 0.367 to 0.711 pm/V, and shortest phase-matching wavelengths ranging from 162.4 to 164.7 nm, highlighting the potential of LiB2O3F as a deep-UV NLO material. These results demonstrate that MLIP-assisted CSP is an effective strategy for discovering new deep-UV NLO materials.